Honestly, it's fucking relatable. A place I worked used to round the time clock to the nearest quarter hour so I would dick around a minute or two so it rolled up instead of down.

They really are coming for our jobs

In some ways yes, but this effect would appear with any kind of reinforcement learning whether it's neural networks or just fuzzy logic. The goal is to promote certain behaviors and if it performs the behaviors that you promoted then the method works.

The problem is that, just like with KPI:s, promoting specific indicators too hard leads to suboptimal results.

It certainly drinks enough water /j

The funniest thing for me is that humans end up doing the exact same thing. This is why it's so notoriously difficult to create organizational policies that actually produce desired results. What happens in practice is that people find ways to comply with the letter of the policy that require the least energy expenditure on their part.

ctrl+f for "calculator", though it doesn't really use the (detailed) wording from the OP, which I think they copied from this list of links without attribution :P

I think this part references it, though it's kinda solely in passing:

Production evaluations can elicit entirely new forms of misalignment before deployment. More importantly, despite being entirely derived from GPT-5 traffic, our evaluation shows the rise of a novel form of model misalignment in GPT-5.1 – dubbed “Calculator Hacking” internally. This behavior arose from a training-time bug that inadvertently rewarded superficial web-tool use, leading the model to use the browser tool as a calculator while behaving as if it had searched. This ultimately constituted the majority of GPT-5.1’s deceptive behaviors at deployment.